Computation and Language (cs.CL)

Abstract: Although dominant in natural language processing, transformer-based models remain challenged by the task of long-sequence processing, because the computational cost of self-attention operations in transformers swells quadratically with the input sequence length. To alleviate the complexity of long-sequence processing, we propose a simple framework to enable the offthe-shelf pre-trained transformers to process much longer sequences, while the computation and memory costs remain growing linearly with the input sequence lengths. More specifically, our method divides each long-sequence input into a batch of chunks, then aligns the interchunk information during the encoding steps, and finally selects the most representative hidden states from the encoder for the decoding process. To extract inter-chunk semantic information, we align the start and end token embeddings among chunks in each encoding transformer block. To learn an effective hidden selection policy, we design a dual updating scheme inspired by reinforcement learning, which regards the decoders of transformers as environments, and the downstream performance metrics as the rewards to evaluate the hidden selection actions. Our empirical results on real-world long-text summarization and reading comprehension tasks demonstrate effective improvements compared to prior longsequence processing baselines.

Abstract: Pre-trained language models (PLMs) contain vast amounts of factual knowledge, but how the knowledge is stored in the parameters remains unclear. This paper delves into the complex task of understanding how factual knowledge is stored in multilingual PLMs, and introduces the Architecture-adapted Multilingual Integrated Gradients method, which successfully localizes knowledge neurons more precisely compared to current methods, and is more universal across various architectures and languages. Moreover, we conduct an in-depth exploration of knowledge neurons, leading to the following two important discoveries: (1) The discovery of Language-Independent Knowledge Neurons, which store factual knowledge in a form that transcends language. We design cross-lingual knowledge editing experiments, demonstrating that the PLMs can accomplish this task based on language-independent neurons; (2) The discovery of Degenerate Knowledge Neurons, a novel type of neuron showing that different knowledge neurons can store the same fact. Its property of functional overlap endows the PLMs with a robust mastery of factual knowledge. We design fact-checking experiments, proving that the degenerate knowledge neurons can help the PLMs to detect wrong facts. Experiments corroborate these findings, shedding light on the mechanisms of factual knowledge storage in multilingual PLMs, and contribute valuable insights to the field. The source code will be made publicly available for further research.

Abstract: The advent of Large Language Models (LLM) has revolutionized the field of natural language processing, enabling significant progress in various applications. One key area of interest is the construction of Knowledge Bases (KB) using these powerful models. Knowledge bases serve as repositories of structured information, facilitating information retrieval and inference tasks. Our paper proposes LLM2KB, a system for constructing knowledge bases using large language models, with a focus on the Llama 2 architecture and the Wikipedia dataset. We perform parameter efficient instruction tuning for Llama-2-13b-chat and StableBeluga-13B by training small injection models that have only 0.05 % of the parameters of the base models using the Low Rank Adaptation (LoRA) technique. These injection models have been trained with prompts that are engineered to utilize Wikipedia page contexts of subject entities fetched using a Dense Passage Retrieval (DPR) algorithm, to answer relevant object entities for a given subject entity and relation. Our best performing model achieved an average F1 score of 0.6185 across 21 relations in the LM-KBC challenge held at the ISWC 2023 conference.

Abstract: Equipped with Chain-of-Thought (CoT), Large language models (LLMs) have shown impressive reasoning ability in various downstream tasks. Even so, suffering from hallucinations and the inability to access external knowledge, LLMs often come with incorrect or unfaithful intermediate reasoning steps, especially in the context of answering knowledge-intensive tasks such as KBQA. To alleviate this issue, we propose a framework called Knowledge-Driven Chain-of-Thought (KD-CoT) to verify and modify reasoning traces in CoT via interaction with external knowledge, and thus overcome the hallucinations and error propagation. Concretely, we formulate the CoT rationale process of LLMs into a structured multi-round QA format. In each round, LLMs interact with a QA system that retrieves external knowledge and produce faithful reasoning traces based on retrieved precise answers. The structured CoT reasoning of LLMs is facilitated by our developed KBQA CoT collection, which serves as in-context learning demonstrations and can also be utilized as feedback augmentation to train a robust retriever. Extensive experiments on WebQSP and ComplexWebQuestion datasets demonstrate the effectiveness of proposed KD-CoT in task-solving reasoning generation, which outperforms the vanilla CoT ICL with an absolute success rate of 8.0% and 5.1%. Furthermore, our proposed feedback-augmented retriever outperforms the state-of-the-art baselines for retrieving knowledge, achieving significant improvement in Hit performance.

Abstract: Recent progress in deep learning and natural language processing has given rise to powerful models that are primarily trained on a cloze-like task and show some evidence of having access to substantial linguistic information, including some constructional knowledge. This groundbreaking discovery presents an exciting opportunity for a synergistic relationship between computational methods and Construction Grammar research. In this chapter, we explore three distinct approaches to the interplay between computational methods and Construction Grammar: (i) computational methods for text analysis, (ii) computational Construction Grammar, and (iii) deep learning models, with a particular focus on language models. We touch upon the first two approaches as a contextual foundation for the use of computational methods before providing an accessible, yet comprehensive overview of deep learning models, which also addresses reservations construction grammarians may have. Additionally, we delve into experiments that explore the emergence of constructionally relevant information within these models while also examining the aspects of Construction Grammar that may pose challenges for these models. This chapter aims to foster collaboration between researchers in the fields of natural language processing and Construction Grammar. By doing so, we hope to pave the way for new insights and advancements in both these fields.

Abstract: We propose a resampling-based approach for assessing keyness in corpus linguistics based on suggestions by Gries (2006, 2022). Traditional approaches based on hypothesis tests (e.g. Likelihood Ratio) model the copora as independent identically distributed samples of tokens. This model does not account for the often observed uneven distribution of occurences of a word across a corpus. When occurences of a word are concentrated in few documents, large values of LLR and similar scores are in fact much more likely than accounted for by the token-by-token sampling model, leading to false positives. We replace the token-by-token sampling model by a model where corpora are samples of documents rather than tokens, which is much closer to the way corpora are actually assembled. We then use a permutation approach to approximate the distribution of a given keyness score under the null hypothesis of equal frequencies and obtain p-values for assessing significance. We do not need any assumption on how the tokens are organized within or across documents, and the approach works with basically *any* keyness score. Hence, appart from obtaining more accurate p-values for scores like LLR, we can also assess significance for e.g. the logratio which has been proposed as a measure of effect size. An efficient implementation of the proposed approach is provided in the `R` package `keyperm` available from github.

Abstract: With the rapid evolution of large language models (LLMs), new and hard-to-predict harmful capabilities are emerging. This requires developers to be able to identify risks through the evaluation of "dangerous capabilities" in order to responsibly deploy LLMs. In this work, we collect the first open-source dataset to evaluate safeguards in LLMs, and deploy safer open-source LLMs at a low cost. Our dataset is curated and filtered to consist only of instructions that responsible language models should not follow. We annotate and assess the responses of six popular LLMs to these instructions. Based on our annotation, we proceed to train several BERT-like classifiers, and find that these small classifiers can achieve results that are comparable with GPT-4 on automatic safety evaluation. Warning: this paper contains example data that may be offensive, harmful, or biased.

Abstract: We propose an unsupervised method to extract keywords and keyphrases from texts based on a pre-trained language model (LM) and Shannon's information maximization. Specifically, our method extracts phrases having the highest conditional entropy under the LM. The resulting set of keyphrases turns out to solve a relevant information-theoretic problem: if provided as side information, it leads to the expected minimal binary code length in compressing the text using the LM and an entropy encoder. Alternately, the resulting set is an approximation via a causal LM to the set of phrases that minimize the entropy of the text when conditioned upon it. Empirically, the method provides results comparable to the most commonly used methods in various keyphrase extraction benchmark challenges.

Abstract: From the perspective of content safety issues, alignment has shown to limit large language models' (LLMs) harmful content generation. This intentional method of reinforcing models to not respond to certain user inputs seem to be present in many modern open-source instruction tuning datasets such as OpenAssistant or Guanaco. We introduce a novel insight to an instruction-tuned model's performance affected by the presence of alignment in supervised fine-tuning dataset. To be specific, we noticed that alignment acts as if it is poisoning the instruction dataset. Experimentally, we demonstrate that aligned answers significantly worsen the performance of the resulting fine-tuned model's on various reasoning benchmarks such as Big Bench (BBH), Massive Multitask Language Understanding (MMLU), Human Eval, and Discrete Reasoning Over Paragraphs (DROP), performing worse than the counterpart tuned without alignment by 4-33%.

Abstract: Complex text is a major barrier for many citizens when accessing public information and knowledge. While often done manually, Text Simplification is a key Natural Language Processing task that aims for reducing the linguistic complexity of a text while preserving the original meaning. Recent advances in Generative Artificial Intelligence (AI) have enabled automatic text simplification both on the lexical and syntactical levels. However, as applications often focus on English, little is understood about the effectiveness of Generative AI techniques on low-resource languages such as Dutch. For this reason, we carry out empirical studies to understand the benefits and limitations of applying generative technologies for text simplification and provide the following outcomes: 1) the design and implementation for a configurable text simplification pipeline that orchestrates state-of-the-art generative text simplification models, domain and reader adaptation, and visualisation modules; 2) insights and lessons learned, showing the strengths of automatic text simplification while exposing the challenges in handling cultural and commonsense knowledge. These outcomes represent a first step in the exploration of Dutch text simplification and shed light on future endeavours both for research and practice.

Abstract: The Natural Language Processing(NLP) community has been using crowd sourcing techniques to create benchmark datasets such as General Language Understanding and Evaluation(GLUE) for training modern Language Models such as BERT. GLUE tasks measure the reliability scores using inter annotator metrics i.e. Cohens Kappa. However, the reliability aspect of LMs has often been overlooked. To counter this problem, we explore a knowledge-guided LM ensembling approach that leverages reinforcement learning to integrate knowledge from ConceptNet and Wikipedia as knowledge graph embeddings. This approach mimics human annotators resorting to external knowledge to compensate for information deficits in the datasets. Across nine GLUE datasets, our research shows that ensembling strengthens reliability and accuracy scores, outperforming state of the art.

Abstract: Large language models (LLMs) are increasingly capable and prevalent, and can be used to produce creative content. The quality of content is influenced by the prompt used, with more specific prompts that incorporate examples generally producing better results. On from this, it could be seen that using instructions written for crowdsourcing tasks (that are specific and include examples to guide workers) could prove effective LLM prompts. To explore this, we used a previous crowdsourcing pipeline that gave examples to people to help them generate a collectively diverse corpus of motivational messages. We then used this same pipeline to generate messages using GPT-4, and compared the collective diversity of messages from: (1) crowd-writers, (2) GPT-4 using the pipeline, and (3 & 4) two baseline GPT-4 prompts. We found that the LLM prompts using the crowdsourcing pipeline caused GPT-4 to produce more diverse messages than the two baseline prompts. We also discuss implications from messages generated by both human writers and LLMs.

Abstract: In Africa, and the world at large, there is an increasing focus on developing Neural Machine Translation (NMT) systems to overcome language barriers. NMT for Low-resource language is particularly compelling as it involves learning with limited labelled data. However, obtaining a well-aligned parallel corpus for low-resource languages can be challenging. The disparity between the technological advancement of a few global languages and the lack of research on NMT for local languages in Chad is striking. End-to-end NMT trials on low-resource Chad languages have not been attempted. Additionally, there is a dearth of online and well-structured data gathering for research in Natural Language Processing, unlike some African languages. However, a guided approach for data gathering can produce bitext data for many Chadian language translation pairs with well-known languages that have ample data. In this project, we created the first sba-Fr Dataset, which is a corpus of Ngambay-to-French translations, and fine-tuned three pre-trained models using this dataset. Our experiments show that the M2M100 model outperforms other models with high BLEU scores on both original and original+synthetic data. The publicly available bitext dataset can be used for research purposes.

Abstract: As research on machine translation moves to translating text beyond the sentence level, it remains unclear how effective automatic evaluation metrics are at scoring longer translations. In this work, we first propose a method for creating paragraph-level data for training and meta-evaluating metrics from existing sentence-level data. Then, we use these new datasets to benchmark existing sentence-level metrics as well as train learned metrics at the paragraph level. Interestingly, our experimental results demonstrate that using sentence-level metrics to score entire paragraphs is equally as effective as using a metric designed to work at the paragraph level. We speculate this result can be attributed to properties of the task of reference-based evaluation as well as limitations of our datasets with respect to capturing all types of phenomena that occur in paragraph-level translations.

Abstract: Open intent detection, a crucial aspect of natural language understanding, involves the identification of previously unseen intents in user-generated text. Despite the progress made in this field, challenges persist in handling new combinations of language components, which is essential for compositional generalization. In this paper, we present a case study exploring the use of ChatGPT as a data augmentation technique to enhance compositional generalization in open intent detection tasks. We begin by discussing the limitations of existing benchmarks in evaluating this problem, highlighting the need for constructing datasets for addressing compositional generalization in open intent detection tasks. By incorporating synthetic data generated by ChatGPT into the training process, we demonstrate that our approach can effectively improve model performance. Rigorous evaluation of multiple benchmarks reveals that our method outperforms existing techniques and significantly enhances open intent detection capabilities. Our findings underscore the potential of large language models like ChatGPT for data augmentation in natural language understanding tasks.